Connect AutoAdy to your AI assistant, then run these prompts in order: audit first, then root-cause the drop, forecast what's about to fatigue, execute the fixes, and finally make the routine autonomous.
A campaign that was printing leads at a comfortable cost per lead doubles its CPL overnight, and the panic move is almost always wrong. You pause the ad that "must be fatigued," launch three new creatives, and a week later you're no better off, because the creative was never the problem. The drop was a tracking gap, or a budget change that retrained delivery, or a competitor flooding your auction. Your dashboard shows you that something changed. It rarely shows you why.
This playbook gets you the why. AutoAdy ships an MCP server, so once you connect it to Claude or ChatGPT, your assistant can run a structured root-cause investigation over weeks of daily data, separate creative fatigue from audience saturation from tracking issues, and then execute the fix: pause the real bleeders, scale what's genuinely working, and set rules so the next drop gets caught before you notice it.
Key takeaways
- "The ad is fatiguing" is the most over-diagnosed cause of a Meta ads drop. Most drops trace back to something else.
diagnose_dropruns several blind investigators over 35 days of daily data, then an adversarial pass that tries to kill each theory, so you get a ranked cause instead of a guess.- It also returns a "saved from mistake" note: the wrong fix it talked you out of, like killing a winner for the wrong reason.
- Forecasting beats reacting.
get_fatigue_forecasttells you which ads will fatigue and how many days you have.- Every fix is reversible and staged. Autonomy is opt-in, with a human veto.
Setup: connect AutoAdy to your assistant (2 minutes)
- In AutoAdy, open Settings → API Keys and create a personal key.
- In Claude, go to Settings → Connectors → Add custom connector and paste the AutoAdy MCP endpoint with your key. ChatGPT supports custom connectors the same way.
- Ask: "Run a full audit of my account and give me the health score." If it returns numbers, you're connected.
Reading and diagnosing your account works on the free plan. The fixes (pausing, scaling, rules, autopilot) are Pro capabilities, and every one of them is reversible.
Use case 1: Get the lay of the land
Before you diagnose a specific drop, get the whole picture in one pass. You want spend, reach, CTR, CPM, frequency, leads, and CPL at the account level, plus your active campaigns and the ad-level detail on the top ones, and a single health score to tell you how much trouble you're actually in.
The prompt:
Run a full audit of my account: account-level spend, impressions, reach,
CTR, CPM, CPC, frequency, leads, and CPL; all active campaigns with their
metrics; and the ad-level breakdown for the top campaigns. Then give me an
account health score out of 100 with the component breakdown (creative
diversity, fatigue risk, budget efficiency, testing velocity, and waste),
and tell me the one component dragging the score down most.
This calls run_full_audit and get_account_health. The health score isn't a vanity number. It points you at the weakest link, and an account bleeding on "waste" needs a different response than one failing on "testing velocity."
Use case 2: Root-cause the drop
This is the prompt that earns the playbook. Don't ask "is my ad fatigued?", because that biases the answer. Describe what you saw and let the assistant investigate without a predetermined conclusion.
The prompt:
My CPL roughly doubled starting around [DATE / "last Tuesday"]. Diagnose
the root cause. Investigate every plausible explanation (creative fatigue,
budget or bid changes that retrained delivery, traffic-quality shifts,
offer or seasonality effects, and conversion-tracking problems) over the
last several weeks of daily data. For each theory, try to disprove it.
Then give me: the ranked causes, the single highest-leverage action to take
this week, and any mistake this analysis just saved me from making.
This calls diagnose_drop. It runs several blind investigators across roughly 35 days of daily data, each looking at one class of cause, then an adversarial pass that actively tries to refute each theory before it survives. It takes 30 to 60 seconds and returns three things that matter: the ranked causes, the one action worth taking this week, and a "saved from mistake" note.
That last one is the quiet hero. The most expensive errors in paid media are confident wrong moves, like pausing a winning creative because frequency looked high when the real issue was a pixel that stopped firing. The diagnosis names the move you were about to make and tells you not to.
Specificity is what changes here. A generic tool says "this ad is fatiguing." This one says "this ad's CPL rose only on one audience segment while it held on another, so exclude the segment instead of pausing the ad." You only get that when a model can hold the whole account's recent history in context at once.
Use case 3: Forecast fatigue before it hits
The best time to deal with creative fatigue is before it shows up in your CPL. Once you've diagnosed the current drop, look forward.
The prompt:
Give me the fatigue forecast for my active ads: which ones are fatiguing,
how many days each has before it degrades, and which need a replacement
queued now. Then break down creative performance by hook, by visual style,
and by messaging angle so I know what to build the replacements around.
This calls get_fatigue_forecast (survival-analysis predictions with days remaining per ad) and analyze_creative_performance. Together they turn fatigue from a fire you fight into a schedule you manage, so you queue the next creative before the current one tips over.
Use case 4: Execute the fixes
A diagnosis you don't act on is just a tidier spreadsheet. Once you know the causes, act, and keep every action reversible.
The prompt:
Based on the diagnosis: pause every ad flagged as a genuine bleeder (show
me the list and the reason for each before you do it), and scale the
campaigns that are genuinely working by 20%. Don't touch anything where the
drop traced back to tracking rather than performance, and flag those for me
to fix at the source instead.
This can call apply_all_kills (or bulk_pause_ads for a specific list) and scale_budget_relative. Notice the guardrail built into the prompt: ads whose problem was a tracking issue should not be paused, because pausing them would kill a working ad over a reporting bug. If an account is bleeding hard and you want one decisive move, there's a single-shot option:
Run emergency recovery: find the bleeders, pause them, and generate
replacement creatives in one pass.
That calls emergency_recovery. Anything it pauses can be re-enabled, and any creative it generates is staged paused until you launch it.
Use case 5: Make the routine autonomous (with a veto)
You shouldn't run this playbook by hand every week. Once you trust the diagnose-and-fix loop, encode the guardrails as rules and let them run, while you keep a veto. Each rule watches one metric, so set up the ones that matter most as separate rules.
The prompt:
Create a rule in plain English: when an ad set's CPL goes above [TARGET]
over the last 3 days, flag it to pause and notify me. Start it in suggest
mode so I approve before anything actually pauses.
Then create a second rule: when an ad's CTR drops below [TARGET] over the
last 7 days, notify me so I can review it.
This calls create_rule once per rule. Each rule evaluates a single metric (CPL, CTR, frequency, spend, and so on) against a threshold over a lookback window, then pauses, scales, or just notifies. Suggest mode is the veto: the rule proposes the action and waits for your yes, and only once you trust it do you move it to auto. Scaling a proven winner is better done deliberately with the scale_budget_relative prompt from Use case 4 than wired into a rule. For hands-free creative replacement, layer on:
Turn on the creative loop for my main campaign: detect fatiguing ads,
generate replacements automatically, but stage them paused so I approve
before anything launches. And enable autopilot on that campaign with
defaults calibrated to its own history.
That chains enable_creative_loop and enable_autopilot. The loop catches fatigue and queues replacements, and autopilot handles the routine optimization. Both are calibrated to this campaign's own history rather than a generic template, and both keep launching gated behind your approval until you decide otherwise.
How the steps stack
| Stage | Tools | What it does |
|---|---|---|
| Audit | run_full_audit, get_account_health | Whole-account picture plus a score that points at the weakest link |
| Root-cause | diagnose_drop | Ranked causes from blind and adversarial investigation, plus the mistake it saved you from |
| Forecast | get_fatigue_forecast, analyze_creative_performance | Which ads fatigue next, how long you have, and what to build |
| Execute | apply_all_kills, scale_budget_relative, emergency_recovery | Reversible fixes: pause real bleeders, scale real winners |
| Autonomy | create_rule, enable_creative_loop, enable_autopilot | The loop runs itself, with a human veto until you remove it |
The steps compound because diagnosis feeds execution and execution feeds automation. The first time through, you run the prompts by hand. By the last stage, the routine catches the next drop before you'd have noticed it, and it still asks before it acts.
What this means
Most of the damage from a Meta ads drop comes from the response, not the drop itself: the paused winner, the three new creatives that fixed nothing, the week of spend chasing a problem that was never creative. A model that can investigate the whole account's recent history, refute its own theories, and name the mistake you were about to make takes a lot of the cost out of being wrong.
Connect AutoAdy to your assistant and run the Use case 2 prompt the next time something breaks, before you pause anything. The diagnosis takes under a minute, and the "saved from mistake" line alone usually pays for the session.